Understanding Public Opinion through Social Media: Summarization, Stance Annotation, Demographic Inference

Frederick Conrad, Michael Schober, Rebecca S. Dolgin, Mao Li, Peilin Chen, Erik Zhou

Keywords: Social media, abstractive summarization, stance annotation, inferring user characteristics, LLMs

TL;DR: We present three features in a new tool that use LLMs to help a user make sense of a large corpus of social media posts to gain insights into public opinion

Submission Type: Non-Archival

Abstract: Social media posts are a promising source of data for insight into the opinions held by members of the public (or at least users of a social media platform), since they can provide near real-time and lower-cost insights than more traditional methods like surveys and focus groups. Additionally, social media data may reveal the opinions of those who would not necessarily agree to participate in surveys or focus groups. However, there are challenges to using social media data for insights into public opinion: (a) the sheer volume far exceeds what a person can read and digest, and (b) they don’t include demographic information, which is central to survey research. However, advances in AI can help address these challenges. We describe how three tools, embedded in a Social Media (SM) Browser, leverage language models to support the use of social media data in public opinion research. The three tools are: summarization (generating textual summaries of posts), stance annotation (e.g., whether a post expresses support or opposition for a proposition or topic), and inferring the demographic characteristics of the user who created each post (e.g., gender, age, education—not directly available within posts or in users’ profiles). These tools can help researchers develop insights about the topics being discussed, the opinions held about those topics, and what kind of users hold those topics, despite the volume of posts and the paucity of information about users.

Submission Number: 4